“The algorithms are already feeding themselves on their own data.”
Generated text and images get posted to the web, then scraped back in as training data. The models start eating their own output. The result is noise tuned to be liked, and it is cheaper to produce than anything a person writes. The open web loses that fight by default.